Open mostafa-razavi opened 4 years ago
ref1="3.790-120.00_4.680-10.00" (normal distribution) ref2="3.820-120.00_4.720-18.00" (normal distribution) ref3="3.840-118.00_4.740-20.00" (normal distribution) ref4="3.780-123.00_4.730-25.00" (uniform distribution)
132 (train: 13, test: 2)
1234 (train: 123, test: 4)
1) When the NN is trained with input samples that are normally distributed, it has a good performance when test data is also normally distributed. If test data is uniformly distributedd, the performance is not good.
2) 1000 data and 2000 data gives similar performance. I need to determine the optimum number of data. 500 data is clearly not enough.
500, 1000, 2000
3) I need to train and test NN using uniform data and see if the performance is acceptable.
4) Including data from a second reference, improves the performance.
There is a strong relationship between Neff and Zstd and between Neff and Ustd, such that having the Neff, these values can be estimated easily using a simple interpolation.
random_type="uniform" ref1="3.790-120.00_4.680-10.00" ref2="3.820-120.00_4.720-18.00" ref3="3.840-118.00_4.740-20.00" ref4="3.780-123.00_4.730-25.00" ref5="3.900-130.00_4.800-10.00" ref6="3.500-100.00_4.400-40.00"
Architecture: 10-10-10
random_type="uniform" ref1="3.790-120.00_4.680-10.00" ref2="3.820-120.00_4.720-18.00" ref3="3.840-118.00_4.740-20.00" ref4="3.780-123.00_4.730-25.00" ref5="3.900-130.00_4.800-10.00" ref6="3.500-100.00_4.400-40.00"
Architecture: 10-10-10
random_type="uniform" ref1="3.790-120.00_4.680-10.00" ref2="3.820-120.00_4.720-18.00" ref3="3.840-118.00_4.740-20.00" ref4="3.780-123.00_4.730-25.00" ref5="3.900-130.00_4.800-10.00" ref6="3.500-100.00_4.400-40.00"
Architecture: 10-10-10
random_type="uniform" ref1="3.790-120.00_4.680-10.00" ref2="3.820-120.00_4.720-18.00" ref3="3.840-118.00_4.740-20.00" ref4="3.780-123.00_4.730-25.00" ref5="3.900-130.00_4.800-10.00" ref6="3.500-100.00_4.400-40.00"
Architecture: 10-10-10
Hypothesis: If I train a NN with 2000 samples that are normally distributed centered around one reference simulation parameter set, the neural network will adequately represent other reference simulation parameter sets as well, at least for Neff.
Test: 1) Obtain ITIC points for select ITIC points 23DMB at five different parameter sets, i.e. sig_CH3, eps_CH3, sig_CH1, and eps_CH1. 2) Select an ITIC point 3) Determine 2000 normally distributed samples around mean (reference parameter set) with standard deviation of 0.015 A for sigma and 3 K for epsilon. 4) Obtain MBAR predictions for 2000 samples of each reference simulation 5) Construct a neural network for each reference simulation with the difference of sig_CH3, eps_CH3, sig_CH1, and eps_CH1 from reference parameter set as input variables and Neff as output variable. 6) Repeat 2-5 procedure for other ITIC points 7) Test each of the five neural networks using the samples of other four MBAR predictions. 8) Plot the results